File size: 3,065 Bytes
44f8169
 
 
 
 
 
 
 
 
 
 
ae7e74e
44f8169
 
 
 
 
 
 
 
 
 
 
38e38d7
44f8169
 
38e38d7
44f8169
 
 
 
 
 
3ff0776
44f8169
 
 
 
 
 
 
c3e078b
44f8169
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3e078b
44f8169
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
---
language:
- de
- fr
- it
- rm
license: cc-by-4.0
task_categories:
- text-classification
task_ids:
- multi-label-classification
pretty_name: Swiss4Arg- Argument Mining on Swiss Federal Voting Brochures 
configs:
- config_name: default
  data_files:
    - split: train
      path: data/default/train.parquet
- config_name: annotations
  data_files:
    - split: train
      path: data/annotations/train.parquet
---

# Swiss4Arg: Argument Mining

A multilingual argument mining dataset built from Swiss federal voting brochures
in **German, French, Italian, and Romansh**.
The dataset covers 53 federal proposals and provides argument component labels
and argumentation structure (relation + target) annotated by 12 human annotators
(3 per language), consolidated via majority voting.

## Dataset Configurations

### `default`

Long format: one row per segment per language (~12k rows).
Each segment appears four times (once per language) with the shared
majority-vote label. 

```python
from datasets import load_dataset
ds = load_dataset("---/swiss-argument-mining")
# filter one language
de = ds["train"].filter(lambda x: x["language"] == "de")
```

**Columns**

| Column | Description |
|---|---|
| `proposal_id` | Federal proposal identifier |
| `language` | `de` / `fr` / `it` / `rm` |
| `side` | `pro` or `con` section of the brochure |
| `id` | Segment ID within document |
| `chapter_type` | `major claim` / `Pro` / `Con` |
| `text` | Segment text in the given language |
| `component` | `Major Claim` / `Claim` / `Premise` / `Non-Argumentative` |
| `relation` | `Support` / `Attack` / `null` |
| `target` | Target segment (e.g. `Claim 4`, `Major Claim`) / `null` |

### `annotations`

Wide format: one row per segment (~3k rows) with all 12 individual
annotator columns alongside the majority-vote labels.
Use this for inter-annotator agreement research or disagreement analysis.

```python
ds = load_dataset("---/swiss-argument-mining", "annotations")
```

**Columns:** `proposal_id`, `side`, `id`, `chapter_type`, `level`,
`text_de/fr/it/rm`, `component_{lang}_{1-3}` × 12,
`relation_{lang}_{1-3}` × 12, `target_{lang}_{1-3}` × 12,
`majority_component`, `majority_relation`, `majority_target`.

## Annotation Schema

Following Stab & Gurevych (2017), the argumentative hierarchy is:

```
Premise ──supports/attacks──▶ Claim ──supports/attacks──▶ Major Claim
```

- **Major Claim**: central thesis of the brochure (one per document, pre-annotated)
- **Claim**: debatable sub-position directly targeting the major claim
- **Premise**: evidence or fact supporting/attacking a specific claim
- **Non-Argumentative**: headers, procedural text, neutral background

## Dataset Statistics

| | |
|---|---|
| Proposals | 53 |
| Languages | German, French, Italian, Romansh |
| Annotators | 12 (3 per language) |
| Total segments | 2,987 |
| Premises | 1,795 (60.1%) |
| Claims | 788 (26.4%) |
| Non-Argumentative | 298 (10.0%) |
| Major Claims | 106 (3.5%) |

## Citation

```bibtex
@article{coming-soon,
  title   = {coming soon},
}
```